A numerical control machine tool part machining data storage method and system

Abstract

The application discloses a kind of numerical control machine tool part machining data storage method and system, a kind of numerical control machine tool part machining data storage method real-time acquisition and synchronous processing process multi-source data;Process section segmentation is carried out based on process entropy to data stream, extract multimodal feature and pass through fusion rule and machine learning's classifier evaluation data value, accordingly adopt hierarchical storage strategy to generate feature data package;For each part binding data gene code, construct machining process knowledge graph, realize whole chain data association and accurate traceability;Sample is generated based on feature data package, process optimization model is trained in cloud using graph neural network and bayesian optimization, and optimization parameter is fed back to machine tool execution, form closed loop, the application adopts process entropy segmentation, knowledge graph and GNN closed loop optimization, solve the problem that data storage cost is high, traceability is difficult, data value is not mined, realize data intelligent compression, whole life cycle traceability and process self-adaptive optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing technology, and in particular to a data storage method and system for CNC machine tool parts processing. Background Technology

[0002] In existing technologies, CNC machine tools are the core production equipment in modern intelligent manufacturing. Their machining processes generate massive amounts of data, including internal CNC system instructions and status data such as G-codes and servo tracking errors, programmable logic controller (PLC) I / O signals such as tool numbers and coolant status, and data from various external sensors such as vibration, acoustic emission, temperature, and power. This data is extremely valuable for process optimization, quality traceability, predictive maintenance, and production management. Currently, the industry faces the following main problems and shortcomings in CNC machining data storage and management:

[0003] 1. Inefficient and costly storage methods: The common practice is to adopt two extreme modes: "full storage" or "result storage". The former records all raw high-frequency data indiscriminately, which leads to the rapid exhaustion of storage space and network bandwidth and a surge in costs. The latter only saves the final processing results, such as size reports, discarding valuable processing information, making process analysis and problem tracing impossible.

[0004] 2. Severe data silos and difficulty in traceability: Process data, material data, and quality inspection data are usually scattered in different information systems, such as MES, ERP, QMS, or databases, lacking an effective correlation mechanism. When a part has a quality problem, it is difficult to quickly and accurately locate the defect to the specific processing step, machine tool status, or even process parameters, resulting in low troubleshooting efficiency.

[0005] 3. Data value has not been fully explored: The stored data is mostly "dormant" archives, lacking effective analysis and utilization methods. It is impossible to automatically learn process rules from historical data, and it is impossible to achieve intelligent and adaptive optimization of process parameters, which limits the further improvement of processing quality, efficiency and consistency.

[0006] Chinese invention patent application number 202211097029.7 discloses a method for data acquisition, storage, and application of CNC machine tools. The method includes a CNC machine tool data acquisition method: installing vibration sensors, noise sensors, alternating current transmitters, and alternating voltage transmitters on the machine tool, and acquiring real-time data from various sensors via serial communication technology. The CNC machine tool data storage method includes: real-time data and information being processed by software and displayed in real-time on the human-machine interface of the industrial computer and the CNC machine tool controller, achieving local data storage in the database of the industrial computer's backend detection and monitoring system; or uploading the data to a network cloud platform server via Ethernet through an IoT smart gateway, achieving cloud storage and cloud computing, with the cloud-stored data displayed through software on a PC or mobile terminal. However, although this method of CNC machine tool data acquisition, storage and application can automatically collect key process indicator data in real time, its storage method is crude, costly, has serious data silos, is difficult to trace, and the data value has not been deeply explored. Therefore, there is an urgent need for a new data storage method and system that can intelligently identify data value, realize data association throughout the entire life cycle, and drive continuous process optimization. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for storing data during the machining of CNC machine tool parts.

[0008] To achieve the above objectives, the technical solution proposed by this invention is as follows:

[0009] A method for storing data on CNC machine tool parts machining, comprising the following steps:

[0010] S1. Real-time acquisition and synchronization: Real-time acquisition of multi-source data streams during the CNC machine tool machining process, and storage of all data in a buffer after being tagged with a unified time stamp; the multi-source data streams include at least internal status data of the CNC system, PLC signals and external sensor data.

[0011] S2. Feature perception and intelligent compression: Real-time analysis of the data stream in the buffer area is performed, and the data is divided into multiple process segments based on process state changes. For each process segment, its multimodal features are extracted and its data value level is evaluated. According to the value level, the corresponding storage strategy is adopted to compress or summarize the original data of the process segment, generating a feature data package containing metadata and feature vectors.

[0012] S3. Full-chain association and traceability index: Bind a unique data gene code to each processed part, associate and bind all feature data packages generated in step S2 with the data gene code of the corresponding part, and jointly construct a traceable data gene chain with the upstream material data and downstream quality inspection data of the part.

[0013] S4. Closed-loop learning and optimization feedback: Based on the feature data package and the associated processing results, generate "process-feature-result" triplet samples, upload high-value samples to the central knowledge base, train the process optimization model using aggregated multi-source samples, distribute the trained optimization model to the edge side for generating optimized process parameter recommendations for new processing tasks, and feed back the new data after execution to the central knowledge base to realize the iterative update of the model.

[0014] The step S2, "dividing it into multiple process segments based on process state changes," specifically includes:

[0015] S21. Initial segmentation is performed based on preset strong segmentation trigger events, including G-code line number change, tool change completion signal, or feed rate returning to zero.

[0016] S22. In the continuous processing interval without strong segmentation triggering events, the process entropy value of key variables is calculated using a sliding window. When the change in the process entropy value exceeds a preset threshold, it is determined that a significant change has occurred in the processing state, and soft segmentation is performed at this point.

[0017] Wherein, the process entropy value The calculation formula is:

[0018]

[0019] in, The total number of joint states for the key variables. Indicates within the time window The first internal observation A kind of joint state, This represents the empirical probability of this state occurring within the window. It is the base of the logarithm.

[0020] Step S2, "extracting its multimodal features," includes extracting frequency domain features, specifically:

[0021] Timing signals of vibration or acoustic emission sensors Perform a fast Fourier transform and calculate its power spectral density. :

[0022]

[0023] in, For the signal length, from the power spectral density The first Q dominant frequency components and their corresponding amplitudes are extracted and used as the frequency domain feature vectors of this process segment.

[0024] The specific steps in step S2, "assessing the data value level," are as follows:

[0025] The extracted multimodal features and semantic features parsed from the G code are input into a predefined value assessment classifier. The value assessment classifier combines a rule base and a lightweight machine learning model to output the value level of the process segment. The value level includes at least three levels: high, medium, and low.

[0026] The rule base contains judgment rules based on machining type, participating axis, and tool information.

[0027] Specifically, this is achieved through a value assessment classifier, whose decision function... Represented as:

[0028]

[0029] in, The input consists of multimodal features and semantic feature vectors. As a value level category, The score is based on a rule matching degree from a preset rule base, which includes judgment rules in the form of "IF machining type == 'fine milling' AND participating axis includes 'C-axis' THEN value level = 'high'". This represents the class probabilities output by a lightweight machine learning model. This is a weighting coefficient used to adjust the weights of the rules and the model output.

[0030] The specific meaning of "adopting the corresponding storage strategy" in step S2 is as follows:

[0031] If the value level is "high", then the original high-frequency data and complete feature set of the process segment are saved using a lossless or near-lossless compression algorithm and stored in a high-speed storage medium.

[0032] If the value level is "medium", then the downsampled data and feature set are saved and stored in a conventional storage medium.

[0033] If the value level is "low", only the statistical features, semantic features and processing result data of the process segment will be saved, and the original data will be discarded.

[0034] The step S3, "constructing a traceable data gene chain," specifically includes:

[0035] S31. Using the data gene code as the primary key, construct a processing knowledge graph that includes entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results.

[0036] S32. Each of the feature data packets is associated and stored as an attribute of its corresponding process segment entity, and is connected to the upstream process, part entity and downstream quality inspection result entity through graph relationships.

[0037] Among them, the knowledge graph of the processing process Represented as:

[0038]

[0039] in, It is a set of nodes, containing entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results. Let be the set of edges, representing the relationships between entities.

[0040] It also includes intelligent traceability steps:

[0041] When a quality anomaly alarm is received for a specific data gene code, the knowledge graph of the processing process is traversed starting from the corresponding part entity to locate all associated process segment entities and their corresponding feature data packets. Suspicious stages are selected based on the process segment attributes, and their stored raw data is retrieved for in-depth analysis.

[0042] The specific steps in step S4, "optimizing the model using aggregated multi-source sample training technology," are as follows:

[0043] Using the knowledge graph of the processing process as the organizational structure of the training data, a graph neural network model is used for training. The input of the graph neural network model is the node features and edge relationships in the graph, and the output is the prediction of the adjustment of process parameters. Its training objective is to minimize the loss between the predicted features / results and the actual features / results.

[0044] Graph neural network models update node representations through message passing mechanisms.

[0045] For each node in the knowledge graph In the Layer representation Calculated using the following formula:

[0046]

[0047] in, Represents a node The set of neighboring nodes, For aggregate functions, For trainable weight matrix, The non-linear activation function is the model's training loss function. This is a combination of mean squared error and cross-entropy loss:

[0048]

[0049] in, and These represent the process characteristics and processing results predicted by the model, respectively. and For the true value, and This is the loss weighting coefficient.

[0050] The "recommendation of optimized process parameters for new processing tasks" in step S4 specifically refers to:

[0051] Based on the part information, cutting tool and machine tool context of the new task, a similar subgraph structure is matched or constructed in the central knowledge base. The trained graph neural network model is used to reason about the subgraph, and the optimal combination of process parameters that makes the predicted machining result is solved by the optimization algorithm. This optimal set of recommended parameters is then sent to the edge side.

[0052] Build a query subgraph based on the new task context. The trained graph neural network model is used as the evaluation function, and a Bayesian optimization algorithm is employed to find the optimal combination of process parameters. :

[0053]

[0054] in, This represents a graph neural network model whose input is the injection process parameters. Subgraph of subsequent queries The output consists of the predicted processing results and stability indices. To map the model output to a function that maps the total utility value, This represents the expected utility under model prediction uncertainty.

[0055] A data storage system for CNC machine tool parts machining, used to implement the above method, the system comprising:

[0056] Edge intelligent storage nodes, deployed on the side of CNC machine tools, include:

[0057] The data acquisition and synchronization module is used to perform step S1.

[0058] A feature extraction and compression engine is used to perform step S2.

[0059] The local storage and management module is used to store feature data packets and associated indexes.

[0060] First communication interface.

[0061] The central knowledge base and optimization platform, deployed in the cloud or on a server, include:

[0062] The second communication interface is used to communicate with the edge intelligent storage node.

[0063] The knowledge graph management module is used to perform the function of constructing and maintaining the knowledge graph of data gene chain and processing process in step S3.

[0064] A machine learning training platform is used to perform the model training and optimization functions in step S4.

[0065] Global repository.

[0066] The edge intelligent storage node interacts with the central knowledge base and optimization platform via a network to collaboratively complete the storage, management and optimization of the CNC machine tool parts processing data.

[0067] The beneficial effects of this invention are:

[0068] 1. Achieve intelligent data compression and significantly reduce storage costs: By analyzing the status of the processing process in real time, data is classified by value and stored in a differentiated manner, effectively reducing the amount of redundant data storage while retaining key process details.

[0069] 2. Construct a full-chain data gene chain to achieve precise traceability: Through unique data gene codes and knowledge graph technology, process data is strongly correlated with upstream and downstream business data, enabling minute-level root cause localization of quality problems.

[0070] 3. Form a closed loop for process optimization and drive continuous improvement: Use stored high-value data to train artificial intelligence models to achieve intelligent recommendation and adaptive optimization of process parameters, forming a virtuous cycle of "data-driven optimization".

[0071] 4. Flexible system architecture, easy to deploy and integrate: The edge-cloud collaborative architecture ensures the real-time performance of data processing and the reliability of the system, while supporting gradual deployment within the existing factory information system. Attached Figure Description

[0072] Figure 1 This is an overall flowchart of the data storage method for CNC machine tool parts processing according to the present invention;

[0073] Figure 2 This is a schematic diagram of the feature perception and intelligent compression process in step S2 of the present invention;

[0074] Figure 3 This is a schematic diagram of the closed-loop learning and optimization feedback process in step S4 of the present invention;

[0075] Figure 4 This is a structural block diagram of the CNC machine tool parts processing data storage system of the present invention. Detailed Implementation

[0076] The present invention will now be described in further detail with reference to the accompanying drawings.

[0077] Example 1:

[0078] A method for storing data on CNC machine tool parts machining, comprising the following steps:

[0079] S1. Real-time acquisition and synchronization: Real-time acquisition of multi-source data streams during the CNC machine tool machining process, and storage of all data in the buffer after being tagged with a unified time stamp; the multi-source data streams include at least the internal status data of the CNC system, PLC signals and external sensor data.

[0080] S2. Feature Awareness and Intelligent Compression: Real-time analysis of the data stream in the buffer area is performed, and the data is divided into multiple process segments based on process state changes. For each process segment, its multimodal features are extracted and its data value level is evaluated. Based on the value level, the corresponding storage strategy is adopted to compress or summarize the raw data of the process segment, generating a feature data package containing metadata and feature vectors.

[0081] S3. Full-chain association and traceability index: Bind a unique data gene code to each processed part, associate and bind all feature data packages generated in step S2 with the data gene code of the corresponding part, and jointly construct a traceable data gene chain with the upstream material data and downstream quality inspection data of the part.

[0082] S4. Closed-loop learning and optimization feedback: Based on the feature data package and the associated processing results, generate "process-feature-result" triple samples, upload high-value samples to the central knowledge base, train the process optimization model using aggregated multi-source samples, distribute the trained optimization model to the edge side for generating optimized process parameter recommendations for new processing tasks, and feed back the new data after execution to the central knowledge base to realize the iterative update of the model.

[0083] Step S2, "dividing it into multiple process segments based on process state changes," specifically includes:

[0084] S21. Initial segmentation is performed based on preset strong segmentation trigger events, including G-code line number change, tool change completion signal, or feed rate returning to zero.

[0085] S22. In the continuous processing interval without strong segmentation trigger events, the process entropy value of key variables is calculated using a sliding window. When the change in the process entropy value exceeds a preset threshold, it is determined that the processing state has changed significantly, and soft segmentation is performed at this point.

[0086] Among them, process entropy value The calculation formula is:

[0087]

[0088] in, The total number of joint states for the key variables. Indicates within the time window The first internal observation A kind of joint state, This represents the empirical probability of this state occurring within the window. It is the base of the logarithm.

[0089] Step S2, "extracting its multimodal features," includes extracting frequency domain features, specifically:

[0090] Timing signals of vibration or acoustic emission sensors Perform a fast Fourier transform and calculate its power spectral density. :

[0091]

[0092] in, For signal length, from power spectral density The first Q dominant frequency components and their corresponding amplitudes are extracted and used as the frequency domain feature vectors of this process segment.

[0093] Step S2, "Assessing the level of data value," specifically involves:

[0094] The extracted multimodal features and semantic features parsed from the G code are input into a predefined value assessment classifier. The value assessment classifier combines a rule base with a lightweight machine learning model to output the value level of the process segment. The value level includes at least three levels: high, medium, and low.

[0095] The rule base contains judgment rules based on machining type, participating axis, and tool information.

[0096] Specifically, this is achieved through a value assessment classifier, whose decision function... Represented as:

[0097]

[0098] in, The input consists of multimodal features and semantic feature vectors. As a value level category, To determine the rule matching score based on a preset rule base, the rule base includes judgment rules in the form of "IF machining type == 'finish milling' AND participating axis includes 'C-axis' THEN value level = 'high'". This represents the class probabilities output by a lightweight machine learning model. This is a weighting coefficient used to adjust the weights of the rules and the model output.

[0099] The specific meaning of "adopting the corresponding storage strategy" in step S2 is as follows:

[0100] If the value level is "high", then the original high-frequency data and complete feature set of the process segment are saved using a lossless or near-lossless compression algorithm and stored in a high-speed storage medium.

[0101] If the value level is "medium", then the downsampled data and feature set are saved and stored in a conventional storage medium.

[0102] If the value level is "low", only the statistical features, semantic features and processing result data of the process segment will be saved, and the original data will be discarded.

[0103] Step S3, "Constructing a traceable data gene chain," specifically includes:

[0104] S31. Using data gene codes as the primary key, construct a machining process knowledge graph that includes entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results.

[0105] S32. Each feature data packet is associated and stored as an attribute of its corresponding process segment entity, and is connected to upstream process, part entity and downstream quality inspection result entity through graph relationships.

[0106] Among them, the knowledge graph of the processing process Represented as:

[0107]

[0108] in, It is a set of nodes, containing entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results. Let be the set of edges, representing the relationships between entities.

[0109] It also includes intelligent traceability steps:

[0110] When a quality anomaly alarm is received for a specific data gene code, the knowledge graph of the processing process is traversed starting from the corresponding part entity to locate all related process segment entities and their corresponding feature data packets. Suspicious stages are selected based on the process segment attributes, and their stored raw data is retrieved for in-depth analysis.

[0111] Step S4, "using aggregated multi-source samples to train the process and optimize the model," specifically refers to:

[0112] Using a knowledge graph of the processing process as the organizational structure for training data, a graph neural network model is used for training. The input of the graph neural network model is the node features and edge relationships in the graph, and the output is the prediction of process parameter adjustment. Its training objective is to minimize the loss between the predicted features / results and the actual features / results.

[0113] Graph neural network models update node representations through message passing mechanisms.

[0114] For each node in the knowledge graph In the Layer representation Calculated using the following formula:

[0115]

[0116] in, Represents a node The set of neighboring nodes, For aggregate functions, For trainable weight matrix, The non-linear activation function is the model's training loss function. This is a combination of mean squared error and cross-entropy loss:

[0117]

[0118] in, and These represent the process characteristics and processing results predicted by the model, respectively. and For the true value, and This is the loss weighting coefficient.

[0119] The specific steps in step S4, "Recommended process parameters for generating optimized processes for new machining tasks," are as follows:

[0120] Based on the part information, tool and machine tool context of the new task, similar subgraph structures are matched or constructed in the central knowledge base. The trained graph neural network model is used to reason about the subgraph, and the optimal combination of process parameters that makes the predicted machining result is solved by the optimization algorithm. This set of recommended parameters is then sent to the edge side.

[0121] Build a query subgraph based on the new task context. The trained graph neural network model is used as the evaluation function, and the Bayesian optimization algorithm is employed to find the optimal combination of process parameters. :

[0122]

[0123] in, This represents a graph neural network model whose input is the injection process parameters. Subgraph of subsequent queries The output consists of the predicted processing results and stability indices. To map the model output to a function that maps the total utility value, This represents the expected utility under model prediction uncertainty.

[0124] Example 2:

[0125] A CNC machine tool parts processing data storage system is provided to implement the parts processing data storage method in Embodiment 1. The system includes:

[0126] Edge intelligent storage nodes, deployed on the side of CNC machine tools, include:

[0127] The data acquisition and synchronization module is used to execute step S1.

[0128] A feature extraction and compression engine is used to perform step S2.

[0129] The local storage and management module is used to store feature data packets and associated indexes.

[0130] First communication interface.

[0131] The central knowledge base and optimization platform, deployed in the cloud or on a server, include:

[0132] The second communication interface is used to communicate with edge intelligent storage nodes.

[0133] The knowledge graph management module is used to perform the function of constructing and maintaining the knowledge graph of data gene chain and processing process in step S3.

[0134] A machine learning training platform is used to perform the model training and optimization functions in step S4.

[0135] Global repository.

[0136] Among them, the edge intelligent storage nodes interact with the central knowledge base and optimization platform through the network to collaboratively complete the storage, management and optimization of CNC machine tool parts processing data.

[0137] Specifically, the hardware and software components of the CNC machine tool parts machining data storage system are as follows:

[0138] Edge intelligent storage node:

[0139] Hardware: It adopts a ruggedized industrial edge computing gateway, equipped with a multi-core CPU, more than 8GB of RAM, more than 256GB of NVMe-SSD, and integrates a variety of industrial network interfaces, such as Ethernet, serial port and data acquisition card slot.

[0140] Software modules:

[0141] Data acquisition and synchronization service: Based on a real-time operating system, such as Preempt-RT Linux, it runs custom drivers and acquisition threads.

[0142] Real-time signal processing and feature extraction engine: Includes algorithm libraries such as FFT, wavelet transform, and entropy calculation, implemented in C++ / Python.

[0143] Local lightweight database: Uses time-series databases such as InfluxDB or embedded relational databases such as SQLite to manage CDP metadata and indexes.

[0144] Communication Proxy: Enables secure communication with cloud platforms using MQTT or HTTPS.

[0145] 2. Central Knowledge Base and Optimization Platform:

[0146] Hardware: Based on cloud server clusters or enterprise private clouds, including compute nodes, storage nodes and high-performance GPU nodes, which are used for model training.

[0147] Software modules:

[0148] Knowledge graph service: Built on Neo4j or JanusGraph graph database, providing interfaces for adding, deleting, modifying, querying and traversing the graph.

[0149] Distributed storage and computing platform: Employs HDFS, object storage such as MinIO, and Spark frameworks for massive sample data storage and batch processing.

[0150] Machine learning platform: Integrates TensorFlow or PyTorch, and encapsulates the pipeline for GNN model training, hyperparameter tuning, model version management and deployment.

[0151] Business Application Interface (API): Provides data traceability and optimization parameter query services for upper-layer systems such as MES and QMS.

[0152] Example 3:

[0153] A method for storing data on CNC machine tool parts machining, comprising the following steps:

[0154] Step 1: Feature perception and intelligent compression based on temporal domain segmentation and process entropy.

[0155] 1. Synchronous acquisition and buffering of multi-source data:

[0156] Data sources: Real-time capture of G-code line numbers, axis commands / actual positions / speeds / currents, spindle speeds / loads from the CNC system bus; tool numbers and coolant switch status from the PLC; synchronous physical signals acquired from external sensors, such as vibration sensors, acoustic emission sensors, and temperature sensors.

[0157] Synchronization and buffering: All data is tagged with a uniform microsecond-level high-precision time stamp and enters a circular memory buffer of fixed duration.

[0158] 2. Time-domain segmentation of the process section:

[0159] Trigger signals: The strong segmentation trigger points are the G-code line number change, the tool change completion signal, and the feed rate returning to zero.

[0160] Process entropy calculation: In the absence of a strong trigger point, a sliding window is used to calculate the multivariate process entropy. Key variables are selected, such as spindle load and vibration energy. The calculation formula is as follows:

[0161]

[0162] in, This is the probability estimate of the joint state of multiple variables within the sliding window at time t.

[0163] Segmentation Decision: hour, This is an empirical threshold that indicates a significant change in the machining state, such as from idle to cutting, or from roughing to finishing; in other words, soft segmentation is performed at this point.

[0164] 3. Multimodal feature extraction:

[0165] For each segmented process section, three types of features are extracted:

[0166] Statistical characteristics: mean, variance, peak value, and kurtosis of data from each channel.

[0167] Frequency domain characteristics: Perform FFT transformation on vibration and sound signals to extract dominant frequencies and frequency band energy.

[0168] Time-frequency domain characteristics: Statistical analysis of wavelet transform coefficients to obtain the characteristics of non-stationary signals.

[0169] At the same time, the G code is parsed to extract the machining type of this segment, such as milling, drilling, tapping, machining surface features, such as plane, curved surface, and semantic features of the shaft involved in the machining.

[0170] 4. Adaptive tiered storage:

[0171] Value assessment and strategy matching: Design an evaluator that combines rules with a lightweight classification model. Example rule base: "IF machining type == 'fine milling' AND participating axis includes 'C-axis' THEN value level = 'high'". The classification model takes the extracted features as input and outputs "high", "medium" and "low" value levels.

[0172] Stored execution:

[0173] High-value segment: Preserve the original data and complete feature set, and store it in a local SSD high-speed storage area using fast and lossless compression such as LZ4.

[0174] Mid-value segment: Save downsampled data and feature sets, such as downsampling from 10kHz to 1kHz, to the local HDD storage area.

[0175] Low-value segment: Only statistical features, semantic features and key results are saved, such as the size measurement values ​​processed in this segment. The original data is discarded after being kept in the buffer for a period of time.

[0176] Generate a feature data package: The storage results of each process segment are encapsulated into a feature data package, whose data structure includes: package ID, start / end time, machine tool ID, value level, feature vector, and original data pointer / data itself.

[0177] The goal of this step is to segment the continuous raw data stream into meaningful process segments, identify their value, and store them differentially.

[0178] Step 2: Embed the full-chain association and traceability index of the "data gene chain".

[0179] 1. Unique Identifier and Binding:

[0180] Each processed part is assigned a data gene code within the system, which can be physically associated with material RFID or QR codes.

[0181] Before processing begins, the data gene code is bound to the processing task sheet and CNC program name through scanning or system docking.

[0182] Each feature data packet generated during the manufacturing process is automatically written into the data gene code of the part.

[0183] 2. Construct a knowledge graph of the processing procedure:

[0184] Entity definition: Define core entities, such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results.

[0185] Relationship establishment: Using processing task sheets and MDC system data, relationships are automatically established. For example: Part - contains → Process - consists of → Feature data packages corresponding to the process segment; Process segment - uses → Tool; Process - output → Quality inspection result.

[0186] Storage and Indexing: This knowledge graph is stored using a graph database. Each feature data package is used as an attribute or external storage pointer of a process segment node in the graph. The corresponding part node can be quickly located through the data gene code, and all related entities can be traversed.

[0187] 3. Intelligent traceability analysis:

[0188] When a quality alarm occurs, such as when the roundness of a certain data gene code part exceeds the tolerance, the traceability analysis engine is activated.

[0189] First, in the graph database, starting from the part node, find the problem node along the edge "part → quality inspection result", and then reverse along the edge "part ← process ← process segment" to locate all relevant process segment nodes.

[0190] The system filters out suspicious stages based on the attributes of process segment nodes, such as the processing type being "precision turning of outer diameter". Then, based on the feature data packet ID associated with the process segment node, it directly retrieves the holographic data stored in the high-value segment, allowing engineers to conduct in-depth spectrum and waveform comparison analysis to quickly pinpoint the root cause.

[0191] This step assigns an identity and association to each feature data packet, building a panoramic view.

[0192] Step 3: Closed-loop learning and optimization feedback based on edge-cloud collaboration.

[0193] 1. Edge-side sample generation and uploading:

[0194] After each process segment is completed, the edge node generates a "process-feature-result" triplet sample.

[0195] Process: Key parameters used in this section, such as rotational speed S, feed rate F, and depth of cut Ap.

[0196] Features: High-order feature vectors extracted from feature data packets.

[0197] Results: Results measured immediately after machining of this section, such as the dimensions of the online probe, or the final quality attributes obtained through downstream detection systems, such as roughness Ra.

[0198] Intelligent screening and uploading: Not all samples are uploaded. Screening strategies are set, such as: a) the top 10% of samples with the best results; b) samples with abnormal process characteristics, such as samples with excessive vibration energy; c) samples of new tools used for the first time; the selected samples and their associated data gene codes and feature data package IDs are packaged and uploaded to the central knowledge base.

[0199] 2. Cloud-based knowledge base construction and model training:

[0200] Knowledge Base: The central knowledge base consists of two core parts: a relational database that stores all uploaded samples and a graph database that stores the complete knowledge graph.

[0201] Training a graph neural network model:

[0202] The knowledge graph of the machining process is used as input. The nodes in the graph, such as process segments and cutting tools, have characteristic attributes, and the edges represent relationships.

[0203] The training objective is to build a graph neural network model that can learn the complex relationships between nodes and edges in a graph and predict the features and results that a node in a specific process segment should produce under a given process input.

[0204] The loss function considers both feature prediction error and result prediction error. Through training, the model can learn implicit knowledge such as "when using a worn tool A to process material B, if you want to maintain low vibration characteristics, you should reduce the rotational speed S by X%".

[0205] 3. Optimize feedback and closed-loop execution:

[0206] When a new part of the same type, that is, a part with the same or similar process requirements, begins to be processed, the edge node sends an optimization request to the cloud, along with the current context, namely the part material, tool information, and machine tool model.

[0207] The cloud-based GNN model constructs a temporary subgraph based on the current context, and through graph reasoning and optimization algorithms, such as gradient-based optimization or Bayesian optimization, it calculates a new set of recommended process parameters that will yield the best prediction results and the most stable features.

[0208] The recommended set is sent to the edge nodes, and after confirmation by the operators or automatic writing into the CNC system for execution, the operators must have CNC system permissions.

[0209] The new data generated in this process is then processed again in steps one and two to generate new samples and knowledge, which are fed back to the cloud for incremental learning of the GNN model, completing the closed loop.

[0210] This step utilizes the structured data generated in the previous two steps to drive continuous process optimization.

[0211] Example 4:

[0212] This part's machining data storage method is applied to the five-axis linkage milling machining of aero-engine impellers.

[0213] This embodiment takes the five-axis precision milling process of a certain type of aero-engine impeller made of Ti-6Al-4V as an example to illustrate the specific implementation process of this method. This process uses a ball end mill, the machining path is complex, and the requirements for vibration and surface quality are extremely high.

[0214] Step S1: Real-time data acquisition and synchronization.

[0215] The hardware components of deploying edge intelligent storage nodes on machine tools include: an industrial bus acquisition card supporting the OPC UA protocol, a multi-channel synchronous data acquisition card for sensors, an industrial-grade computing unit, such as a high-performance industrial computer and a local solid-state drive, such as an SSD, and a data acquisition and synchronization module running on the software.

[0216] Data collected from the CNC system includes: G-code line number, actual and commanded positions of the X / Y / Z / A / C axes (sampling rate 1kHz), and actual spindle speed and load (500Hz).

[0217] Data collected from the PLC: current tool number, i.e., event trigger and coolant pressure switch status.

[0218] Data collected from sensors includes: triaxial vibration acceleration signal at the spindle housing (sampling rate 20kHz), acoustic emission signal near the spindle end (sampling rate 1MHz), and temperature field data from an infrared thermal imager in the working area (10Hz).

[0219] Synchronization mechanism: All acquisition threads are driven by the same high-precision hardware clock, such as PTP synchronization clock, which gives each data point a uniform microsecond-level timestamp and temporarily stores it in a memory circular buffer.

[0220] Step S2: Feature perception and intelligent compression.

[0221] The edge node feature extraction and compression engine processes buffer data in real time.

[0222] Process section segmentation:

[0223] Strong segmentation: When the system detects that the G code line number jumps from N100 to N101, that is, from finish milling the blade basin surface to finish milling the blade back surface, or receives a PLC signal that the tool change is complete, a new process segment is immediately created.

[0224] Soft segmentation is achieved through a core algorithm: during the continuous milling of a curved surface, the system uses a 1-second sliding window to calculate the joint state entropy of two key variables: spindle load and effective value of Z-axis vibration within the window. For example, if the load and vibration are quantified into three levels—low, medium, and high—there are nine possible joint states.

[0225] Within the time window t, the frequency of occurrence of each state is calculated as an empirical probability. Substituting into the entropy formula:

[0226]

[0227] When the entropy change between two adjacent windows When the value exceeds the preset threshold, it indicates a significant change in processing stability, such as slight tool scratches. At this point, soft segmentation is performed, ending the current process segment and starting a new one. The impeller precision milling is divided into multiple stable process segments S1, S2... and several process segments A1 where minor anomalies are detected.

[0228] Multimodal feature extraction:

[0229] For vibration signals, perform FFT on the data within each process segment. For example, for a 2-second vibration signal... Calculation with N=40000 points The first five peak frequencies (Q=5) in the power spectrum, such as 800Hz, 1200Hz, 2000Hz, etc., and their amplitudes are extracted to form a 10-dimensional frequency domain feature vector, which contains 5 frequency values ​​and 5 amplitudes.

[0230] Simultaneously, time-domain statistical features are extracted, such as mean, variance, kurtosis, and semantic features parsed from G-code, such as "five-axis linkage fine milling", "tool radius R5", and "feed direction along surface normal".

[0231] Value assessment and tiered storage:

[0232] Input the feature vector X into the value assessment classifier. There is a rule in the rule base R: "IF machining type == 'fine milling' AND participating axes include 'A axis' and 'C axis' THEN value level = 'high'". Since the current process is five-axis fine milling of the impeller, the matching score of this rule is R(X, 'high') = 1.0.

[0233] Meanwhile, a pre-trained gradient boosting tree model M, based on patterns learned from historical data, outputs probabilities for feature X: M(X,'High')=0.85, M(X,'Medium')=0.15.

[0234] Let the tradeoff coefficient α = 0.3, then the decision function is calculated.

[0235] .

[0236] The final value level was determined to be "high".

[0237] Storage execution: For the "high-value" impeller milling process section, the LZ4 lossless compression algorithm is used to save all the original 20kHz vibration, 1MHz acoustic emission data and all features, and store them on the local SSD. At the same time, a feature data package, namely CDP, is generated, whose metadata includes: CDP_ID: TURBINE_WHL_FINISH_001, timestamp, machine ID, data gene code: WHL_20231027_001, value level: high, feature vector: [800, 0.5, 1200, 0.3, ...].

[0238] Step S3: Full-chain association and traceability index.

[0239] Data Gene Code Binding: The impeller blank is assigned a QR code upon entering the warehouse, coded as WHL_20231027_001. Before processing, scanning this QR code binds the data gene code DGC to the current processing task.

[0240] Building a knowledge graph:

[0241] The knowledge graph management module of the central knowledge base receives CDP and DGC uploaded by edge nodes. It uses DGC as the primary key to create or update a graph. .

[0242] Create a node: the part is WHL_20231027_001, the process is five-axis finish milling, the machine tool is five-axis milling machine 01, the tool is ball end mill R5-number T123, and create a corresponding process segment node for each CDP, such as process segment (TURBINE_WHL_FINISH_001).

[0243] Create relationship edges: Part - Owns → Operation, Operation - Contains → Process Segment, Process Segment - Uses → Tool, Process Segment - Originates from → Machine Tool.

[0244] Intelligent traceability:

[0245] Subsequent testing revealed that the profile of one of the impeller blades was out of tolerance. The problem was described by inputting DGC-WHL_20231027_001 and a quality issue description into the interactive application layer.

[0246] The tracing engine starts with part WHL_20231027_001 in the graph and traverses to the associated processes, namely five-axis precision milling and all process segment nodes below it. Based on the node attributes, such as "machining area: third blade", it filters out suspicious process segment nodes.

[0247] Based on the CDP_ID associated with these nodes, the system directly retrieves the corresponding holographic vibration data from the edge SSD or central storage. Engineers analyzed the data and found that in a certain process segment CDP marked as "high" value, there was an abnormal peak in the vibration spectrum at 2000Hz. Combined with the tool nodes associated in the spectrum, it was determined that tool T123 had experienced slight wear and chipping in this area, thus accurately locating the root cause.

[0248] Step S4: Closed-loop learning and optimization feedback.

[0249] Sample generation and uploading: For each "high" value impeller milling process segment, edge nodes generate triplet samples.

[0250] Process = {Spindle speed S: 8000 rpm, Feed rate F: 1200 mm / min, Cutting width Ae: 0.2 mm};

[0251] Characteristics = {Dominant vibration frequency: 1200Hz, Amplitude: 0.3g};

[0252] Result = {Subsequent coordinate measuring machine measurement of surface roughness Ra in this area: 0.4 μm}.

[0253] This sample, along with DGC and CDP_ID, was marked as the "best sample" and uploaded.

[0254] Model training:

[0255] The cloud-based machine learning training platform aggregates thousands of impeller samples processed by multiple machine tools.

[0256] It uses a knowledge graph G as its structure to train a graph neural network GNN. The AGGREGATE function of the GNN adopts an attention mechanism. For example, for a process segment node, it aggregates the wear state features of the tool nodes connected to it and the stiffness features of the machine tool nodes.

[0257] The model uses a message passing formula:

[0258]

[0259] The node characterization is updated to ultimately predict the process adjustment amount. The loss function L is optimized for both vibration characteristics and surface roughness, that is, to improve the prediction accuracy of MSE loss and CE loss.

[0260] Optimization Recommendation:

[0261] When a new batch of impellers of the same model, DGC: WHL_20231105_XXX, begins processing, the system matches similar graph substructures in the knowledge base based on the blank batch and tool information associated with its DGC. .

[0262] The current plan parameters injection The process segment nodes utilize a pre-trained GNN model. The Bayesian optimizer uses the predicted results as its evaluation function to search within the parameter space, such as S∈[7500,8500], F∈[1100,1300], for the function that maximizes the expected utility. parameters The utility function U combines the predicted Ra value and vibration stability. The smaller the Ra value, the better, and the more stable the vibration stability, the better.

[0263] Ultimately, the system recommends a set of optimization parameters. The data is sent to the edge nodes to guide the current processing. The new data generated after processing re-enters the system loop and is used for incremental training of the GNN model.

[0264] Example 5:

[0265] This part's machining data storage method is applied to the precision turning of automotive gearbox shafts.

[0266] This embodiment takes the precision turning process of an automotive gearbox output shaft as an example to highlight the application of this method in the machining of shaft parts and the differences from Embodiment 4.

[0267] Step S1 is the same as step S1 in Example 4.

[0268] Step S2: Feature perception difference:

[0269] Segmentation: Strong segmentation trigger events are more frequent, including "feed rate zeroing" for measurement points and "G-code switching to thread cutting command".

[0270] Feature extraction: Frequency domain analysis focuses on monitoring chatter frequencies during turning, performing FFT on the acoustic emission signal, and analyzing the power spectrum. Energy in the 2-5 kHz range was used as a key characteristic to evaluate cutting stability.

[0271] Value Assessment: The rule base contains: "IF machining feature == 'thread turning' THEN value level = 'high'" and "IF machining feature == 'rough turning' THEN value level = 'medium'". For the rough turning segment, the GBDT model may output a "low" probability due to stable load. The final decision is then processed by the decision function. The overall assessment is "medium".

[0272] Storage strategy: For the "medium" level coarse travel segment, the original current signal is downsampled from 2kHz to 500Hz and then saved; for the "low" level idle travel segment, only the start and end times, energy consumption, and travel distance are saved.

[0273] Step S3: Knowledge Graph Construction

[0274] The construction of the graph G is more hierarchical. Under a part GEAR_SHAFT_XXX node, multiple process nodes are associated, such as rough turning, finish turning of the outer circle, milling of keyway, and grinding of the outer circle. Each process node contains multiple process segment nodes. The process segment node of the grinding of the outer circle process will be associated with the online measuring instrument node and the real-time dimension detection result node provided by the detection equipment used, forming a tighter machining-inspection closed loop chain.

[0275] Key points of closed-loop optimization in step S4:

[0276] For turning, the optimization objective focuses more on dimensional accuracy, such as diameter tolerance and shape accuracy, such as roundness. The GNN model learns how to compensate for tool wear by adjusting the spindle speed (S) and feed rate (F), which is reflected in tool node features and machine tool thermal deformation, which is reflected in the time-series temperature features of machine tool nodes. When the model detects increased tool wear, it optimizes the recommended parameters using Bayesian methods. It may actively reduce the feed F to maintain dimensional stability.

[0277] Working principle:

[0278] The system adopts a three-layer architecture: an edge intelligence layer, a central knowledge layer, and an interactive application layer. The edge intelligence layer is responsible for feature extraction, compression storage, and preliminary analysis of real-time data; the central knowledge layer is responsible for data aggregation, correlation modeling, and deep learning; and the interactive application layer provides human-machine interfaces such as quality traceability and process optimization recommendations. Its workflow is a process of continuous refinement and enhancement of data value: the raw data is first refined into high-density feature data packages through feature perception and intelligent compression; these data packages are given complete contextual semantics through full-chain association and traceability indexing, forming traceable knowledge units; finally, these knowledge units are transformed into optimization parameters that drive process improvement through closed-loop learning and optimization feedback, forming a continuously evolving intelligent cycle.

[0279] The beneficial effects of this invention are: achieving intelligent data compression and significantly reducing storage costs; by analyzing the processing status in real time, data is classified and stored differently based on its value, effectively reducing redundant data storage while retaining key process details; constructing a full-link data gene chain to achieve precise traceability; using unique data gene codes and knowledge graph technology to strongly correlate processing data with upstream and downstream business data, enabling minute-level root cause localization of quality problems; forming a closed loop for process optimization to drive continuous improvement; using stored high-value data to train artificial intelligence models to achieve intelligent recommendation and adaptive optimization of process parameters, forming a virtuous cycle of "data-driven optimization"; and a flexible system architecture that is easy to deploy and integrate; the edge-cloud collaborative architecture ensures the real-time performance of data processing and the reliability of the system, while supporting gradual deployment within existing factory information systems.

[0280] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for storing data during the machining of CNC machine tool parts, characterized in that, Includes the following steps: S1. Real-time acquisition and synchronization: Real-time acquisition of multi-source data streams during the CNC machine tool machining process, and storage of all data in a buffer after being stamped with a unified time stamp; the multi-source data streams include at least internal status data of the CNC system, PLC signals and external sensor data. S2. Feature perception and intelligent compression: Real-time analysis of the data stream in the buffer area, dividing it into multiple process segments based on process state changes, extracting multimodal features for each process segment and evaluating its data value level, and using the corresponding storage strategy to compress or summarize the original data of the process segment according to the value level, generating a feature data package containing metadata and feature vectors. S3. Full-chain association and traceability index: Bind a unique data gene code to each processed part, associate and bind all feature data packages generated in step S2 with the data gene code of the corresponding part, and jointly construct a traceable data gene chain with the upstream material data and downstream quality inspection data of the part. S4. Closed-loop learning and optimization feedback: Based on the feature data package and the associated processing results, generate "process-feature-result" triplet samples, upload high-value samples to the central knowledge base, train the process optimization model using aggregated multi-source samples, distribute the trained optimization model to the edge side for generating optimized process parameter recommendations for new processing tasks, and feed back the new data after execution to the central knowledge base to realize the iterative update of the model. Step S2, "dividing it into multiple process segments based on process state changes," specifically includes: S21. Initial segmentation is performed based on preset strong segmentation trigger events, including G-code line number change, tool change completion signal, or feed rate returning to zero; S22. In the continuous processing interval without strong segmentation triggering events, the process entropy value of key variables is calculated using a sliding window. When the change in the process entropy value exceeds a preset threshold, it is determined that a significant change has occurred in the processing state, and soft segmentation is performed at this point. Wherein, the process entropy value The calculation formula is: in, The total number of joint states for the key variables. Indicates within the time window The first internal observation A kind of joint state, This represents the empirical probability of this state occurring within the window. It is the base of the logarithm; The specific meaning of "adopting the corresponding storage strategy" in step S2 is as follows: If the value level is "high", then the original high-frequency data and complete feature set of the process segment are saved using a lossless or near-lossless compression algorithm and stored in a high-speed storage medium. If the value level is "medium", then save the downsampled data and feature set to a regular storage medium; If the value level is "low", only the statistical features, semantic features and processing result data of that process segment will be saved, and the original data will be discarded.

2. The data storage method for CNC machine tool parts processing as described in claim 1, characterized in that, The step S2, "extracting its multimodal features," includes extracting frequency domain features, specifically: Timing signals of vibration or acoustic emission sensors Perform a fast Fourier transform and calculate its power spectral density. : in, For the signal length, from the power spectral density The first Q dominant frequency components and their corresponding amplitudes are extracted and used as the frequency domain feature vectors of this process segment.

3. The data storage method for CNC machine tool parts processing as described in claim 1, characterized in that, The specific steps in step S2, "assessing the data value level," are as follows: The extracted multimodal features and semantic features parsed from the G code are input into a predefined value assessment classifier. The value assessment classifier combines a rule base and a lightweight machine learning model to output the value level of the process segment. The value level includes at least three levels: high, medium, and low. The rule base includes judgment rules based on machining type, participating axis, and tool information; Specifically, this is achieved through a value assessment classifier, whose decision function... Represented as: in, The input consists of multimodal features and semantic feature vectors. As a value level category, The score is based on a rule matching score from a preset rule base, which includes judgment rules in the form of "IF machining type == 'fine milling' AND participating axis includes 'C-axis' THEN value level = 'high'". This represents the class probabilities output by a lightweight machine learning model. This is a weighting coefficient used to adjust the weights of the rules and the model output.

4. The data storage method for CNC machine tool parts processing as described in claim 1, characterized in that, The step S3, "constructing a traceable data gene chain," specifically includes: S31. Using the data gene code as the primary key, construct a processing knowledge graph that includes entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results; S32. Each of the feature data packets is associated and stored as an attribute of its corresponding process segment entity, and is connected to the upstream process, part entity and downstream quality inspection result entity through graph relationships; Among them, the knowledge graph of the processing process Represented as: in, It is a set of nodes, containing entities such as parts, processes, process segments, machine tools, cutting tools, and quality inspection results. Let be the set of edges, representing the relationships between entities.

5. The data storage method for CNC machine tool parts processing as described in claim 4, characterized in that, It also includes intelligent traceability steps: When a quality anomaly alarm is received for a specific data gene code, the knowledge graph of the processing process is traversed starting from the corresponding part entity to locate all associated process segment entities and their corresponding feature data packets. Suspicious stages are selected based on the process segment attributes, and their stored raw data is retrieved for in-depth analysis.

6. The data storage method for CNC machine tool parts processing as described in claim 1, characterized in that, In step S4, "optimizing the model using aggregated multi-source sample training process" specifically refers to: Using a knowledge graph of the processing process as the organizational structure of the training data, a graph neural network model is used for training. The input of the graph neural network model is the node features and edge relationships in the graph, and the output is the prediction of the adjustment of process parameters. Its training objective is to minimize the loss between the predicted features / results and the actual features / results. Graph neural network models update node representations through message passing mechanisms: For each node in the knowledge graph In the Layer representation Calculated using the following formula: in, Represents a node The set of neighboring nodes, It is an aggregate function. For trainable weight matrix, The non-linear activation function is the model's training loss function. This is a combination of mean squared error and cross-entropy loss: in, and These represent the process characteristics and processing results predicted by the model, respectively. and For the true value, and This is the loss weighting coefficient.

7. The data storage method for CNC machine tool parts machining as described in claim 6, characterized in that, The "recommendation of optimized process parameters for new processing tasks" in step S4 specifically refers to: Based on the part information, cutting tool and machine tool context of the new task, a similar subgraph structure is matched or constructed in the central knowledge base. The trained graph neural network model is used to reason about the subgraph, and the optimal combination of process parameters that makes the predicted processing result is solved by the optimization algorithm. This set of recommended parameters is then sent to the edge side. Build a query subgraph based on the new task context. The trained graph neural network model is used as the evaluation function, and a Bayesian optimization algorithm is employed to find the optimal combination of process parameters. : in, This represents a graph neural network model whose input is the injection process parameters. Subgraph of subsequent queries The output consists of the predicted processing results and stability indices. To map the model output to a function that maps the total utility value, E ⋅ This represents the expected utility under model prediction uncertainty.

8. A data storage system for machining parts in a CNC machine tool, used to implement the method according to any one of claims 1-7, characterized in that, The system includes: Edge intelligent storage nodes, deployed on the side of CNC machine tools, include: The data acquisition and synchronization module is used to perform step S1; Feature extraction and compression engine, used to perform step S2; The local storage and management module is used to store feature data packets and associated indexes; First communication interface; The central knowledge base and optimization platform, deployed in the cloud or on a server, include: The second communication interface is used to communicate with the edge intelligent storage node; The knowledge graph management module is used to perform the function of constructing and maintaining the knowledge graph of data gene chain and processing process in step S3; A machine learning training platform, used to perform the model training and optimization functions in step S4; Global repository; The edge intelligent storage node interacts with the central knowledge base and optimization platform via a network to collaboratively complete the storage, management and optimization of the CNC machine tool parts processing data.

Images